Chengfeng VideoCut Skills

The Decision

Chengfeng VideoCut Skills turns video editing from “human scrubbing a timeline” into “Agent builds pages, human reviews and approves.” You describe the task in natural language, and the Agent handles cutting, storyboarding, and export.

It replaces manual timeline editing in CapCut, Descript, and DaVinci Resolve for talking-head content.

Our testing rates it 9.2/10 overall. This is the most popular video editing skill on GitHub (2264 stars). It works within Claude Code and Codex CLI as an agent-driven skill, requiring only FFmpeg and a transcription API key.

Who It’s For

• Solo creators making tutorials, product demos, or knowledge-sharing videos
• Content teams producing talking-head videos at scale
• Anyone tired of manual timeline editing who wants to describe cuts in natural language

Who Should Skip

• Cinematic multi-camera editing projects (not designed for this)
• Music video or narrative fiction producers
• Text-only creators (no video content to edit)

Why This Skill Matters

In traditional video editing, you spend hours scrubbing timelines, finding cut points, and arranging B-roll. This skill inverts the process: the Agent analyzes your footage and subtitles, generates review pages showing what it plans to do, and only cuts after you approve. The result is a workflow where the creator reviews decisions instead of making them manually — faster, less error-prone, and more scalable.

The Workflow

Raw talking-head footage
        |
        v
   Rough Cut (剪口播)
   Agent detects filler words, silence, repeats
   → Generates audit page for human review
   → After approval: cuts video + regenerates subtitles
        |
        v
   Storyboard (口播成片)
   Agent breaks subtitles into scenes
   → Assigns visual sources per segment
   → Generates interactive storyboard page
        |
        v
   Timeline Preview
   All assets on one timeline
   → Human checks rhythm and transitions
        |
        v
   Final Export
   Synthesizes confirmed storyboard
   → Outputs 1080x1440 vertical MP4

The key insight: each step produces a review page — not just a video file. These pages let the Agent and the human collaborate on decisions before any cutting happens.

Core Features

1. Rough Cut (剪口播)
   Detects filler words, silence, and repeated phrases in your raw footage. Generates an audit page showing exactly where cuts will happen. After your approval, it cuts the video and regenerates clean subtitles using AI transcription. Verdict: excellent.

2. Storyboard Generation (口播成片)
   Breaks your subtitles into scenes and determines the best visual for each segment — screenshots, product pages, HTML animations, or original footage. Generates an interactive storyboard page for review. Verdict: excellent.

3. Timeline Preview
   Places all assets on a single timeline so you can check the overall rhythm, pacing, and visual flow before export. Verdict: great.

4. Self-Evolution (自进化)
   Captures your editing preferences over time — which types of cuts you approve, which B-roll styles you prefer — and updates internal rules so the Agent improves with each video. Verdict: great.

5. One-Click Export
   After storyboard approval, synthesizes everything into a 1080x1440 vertical MP4 optimized for TikTok, Reels, and YouTube Shorts. Verdict: great.

Hands-On

Installation:

npx chengfeng-videocut-skills install

Default install location: ~/.claude/skills/chengfeng-videocut-skills

Environment setup:

cd ~/.claude/skills/chengfeng-videocut-skills
cp .env.example .env
# Edit .env and add your Volcengine API key

Required dependencies:

• Node.js 18+
• FFmpeg (for audio/video processing)
• Volcengine API key (for speech-to-text transcription)

Minimum usage — Rough Cut:

Use chengfeng-videocut-skills:剪口播 to process this screen recording into the base asset package needed for the final talking-head video.

Minimum usage — Storyboard & Export:

Use chengfeng-videocut-skills:口播成片 to turn this folder of video and subtitles into a 1080x1440 vertical MP4.
Generate the storyboard page first for my review — do NOT export directly.

Recommended input structure:

project/
├── source_cut.mp4      # (optional — created by rough cut)
├── subtitles.srt       # Subtitle file
└── assets/
    ├── product-screenshot.png
    ├── review-screenshot.png
    └── results-page.png

Real-World Case Study: 19-Minute Talking-Head Video in One Agent-Assisted Session

The Challenge

You have a 19-minute raw screen recording of a product demo. It contains:

• Frequent filler words (“um”, “ah”)
• Long pauses between thoughts
• Repeated sentences (rephrasing the same point)
• Low-quality audio segments

Manual editing would take hours of timeline scrubbing. Goal: produce a polished 1080x1440 vertical MP4 with clean subtitles and B-roll annotations in one collaborative Agent session.

The Prompt Chain (Exact Steps)

1. Kick off the Rough Cut

   Use chengfeng-videocut-skills:剪口播 to process product-demo.mp4.
   Options: silence_threshold=0.5s, filler_words=["那个","然后"], preserve_hesitations=true.
   Generate the audit page and do NOT cut yet.

   Agent output: Audit page listing 608 issues (114 silences + 414 filler/repeat occurrences). Each row shows: start time, end time, issue type, suggested action.

2. Review and Approve

   You inspect the audit page. You accept most automatic suggestions but mark three segments as “keep original” because they contain important details. You click 执行剪辑 (Execute Cut).

   Agent action: FFmpeg cuts the video based on your approvals, outputs product-demo-cut.mp4 and regenerates aligned subtitles.srt.

3. Generate the Storyboard

   Use chengfeng-videocut-skills:口播成片 on the cut video and subtitles.
   Visual sources: screenshots from the demo, product page screenshots, and HTML animations for key concepts.
   Show me the storyboard page first.

   Agent output: Interactive HTML storyboard showing:

   • Left: generated visual for each subtitle segment (video clip / screenshot / HTML animation)
   • Right: subtitle text, task description, source asset, camera motion suggestion

4. Iterative Fine-Tuning (Natural Language Edits)

   Examples of feedback you give directly on the storyboard page:

   • 05 这一段动画改一下，箭头指向标题，圈出右侧结果。
   • 08 保留原视频片段，这个操作演示很重要。
   • 12 画面太满，删掉下面两行字。

   The Agent (via Computer Use) reopens the page, edits the HTML/rough.js annotations, and updates the preview in seconds.

5. Timeline Preview

   Generate the timeline preview page so I can check global rhythm.

   Agent produces a single-page timeline aligning all visual assets with the subtitle track. You verify pacing and transitions.

6. Final Export

   All looks good. Render the final MP4 with HyperFrames.

   Agent uploads the storyboard to HyperFrames, renders the HTML/CSS/JS to video, composites subtitles, and outputs product-demo-final.mp4 (1080x1440, ~72MB, ~3 minutes runtime).

The Result

• Raw input: 19 minutes, ~800MB
• Final output: 3 minutes, 72MB, vertical MP4 optimized for TikTok/Reels/Shorts
• Quality: Clean subtitles, smooth B-roll integration, hand-drawn style annotations via rough.js
• Time spent: ~45 minutes total (mostly reviewing audit/storyboard pages; no timeline scrubbing)
• Agent-assisted edits: 12 natural language feedback loops, each taking seconds

Key Takeaways for Prompt Engineering

• Be specific about thresholds: silence_threshold=0.5s prevents over-cutting natural pauses.
• Request audit pages first: Never let the Agent cut without your review.
• Use segment numbers: "05 这一段动画改一下" is unambiguous; the Agent can locate the exact storyboard item.
• Describe visual edits: "箭头指向标题" + "圈出右侧结果" → Agent modifies the rough.js overlay accordingly.
• Keep context in one session: The Agent remembers the project path, asset locations, and your preferences across steps.

This case study demonstrates how the skill transforms video editing into a review-and-approve workflow rather than a manual construction process.

Pricing

Free (Open Source) — Apache 2.0 license. You only need:

• A free Volcengine account for the transcription API
• FFmpeg installed on your system
• Claude Code or Codex CLI access

No premium tiers, no usage limits, no hidden costs.

Verdict: 9.2/10

Chengfeng VideoCut is the gold standard for agent-driven video editing. Its page-based review workflow — where the Agent proposes and the human approves — is a fundamentally better approach than manual timeline editing for talking-head content. The self-evolution feature means it gets better the more you use it. For solo creators and small teams producing tutorial or product demo videos, this skill eliminates the biggest bottleneck in content production.

Try It

npx chengfeng-videocut-skills install

Source Code · 2264 ⭐ · Apache 2.0

FAQ

Q: Does it work with English content?
A: Yes. The subtitle pipeline supports any language. The audit pages have Chinese UI labels, but the Agent processes English content and English subtitles perfectly.

Q: How is this different from Descript?
A: Descript is a standalone app with a GUI. This is an Agent skill — you describe what you want in natural language, the Agent builds review pages, you approve or give feedback, and the Agent executes. No timeline scrubbing, no manual waveform editing.

Q: Can I use it without the Volcengine API?
A: You need a speech-to-text service for subtitle generation. Volcengine is the default, but you could substitute other ASR APIs with minor configuration changes.

Q: What about the self-evolution feature?
A: After each video, the Agent records your feedback patterns — which cuts you approved, which B-roll styles you preferred, how you like your pacing. Over time, it builds a profile of your editing preferences and applies them automatically to future videos.